View
3
Download
0
Category
Preview:
Citation preview
2
INDIRECT ADAPTIVE CONTROL
OUTLINE
1. Introduction a. Main properties b. Running example
2. Adaptive parameter estimation a. Parameterized system model b. Linear parametric model c. Normalized gradient algorithm d. Normalized least-squares algorithm e. Discrete-time version of adaptive algorithms
3. Identification and robustness a. Parametric convergence and persistency of excitation b. Robustness of adaptive algorithms
4. Indirect adaptive control a. Model reference control b. Pole placement control
5. Adaptive observers
3
1. INTRODUCTION Dynamic systems are characterized by their structures and parameters:
Linear:
( ) ( ) ;: ,l= + +Σ = + +
x A θ x B θ u dy Cx Du v
Nonlinear:
( , , , );: ( , ) ,n=Σ = +
x f x u d θy h x u v
x is state vector, u is control input, d is disturbance, y is output, v is noise, θ is parameters.
Control system design steps:
P u y
Pmu y
, m l n∈ Σ ΣP
Pm
u y
Δ uncertainty
C
P
u y C
1. Modeling
2. Control design
3. Implemen-tation
Stability, robustness, performance???
4
a. Main properties
Parameter estimation is to use a collection of available system signals y and u , based on certain system structure information lΣ or nΣ , to produce estimates ( )tθ of the system parameters θ ⇒ Appears on the step 1. Adaptive parameter estimation is a dynamic estimation procedure that produce updated parameter estimates on-line ⇒ Appears on the step 2&3. Adaptive parameter estimation is crucial for indirect adaptive control design where controller parameters ( )c tθ are some continuous functions of the estimates ( )tθ :
Pm(θ)
u y C
Strategy for control adjustment
Pm(θ)
u y C(θс)
Adaptive parameter estimation
Control parameter derivation
( )tθ
θc(t)
The general scheme of adaptive control. The scheme of indirect adaptive control.
5
Key issues in the classical adaptive parameter estimation:
− linear parameterization of system models,
− linear representation of parametric error models,
− stable design of adaptive estimation algorithms,
− analytical proof of system stability,
− parameter convergence,
− robustness of adaptive estimation.
Realization:
− continuous-time,
− discrete-time.
6
b. Running example
Moving vehicle:
Ff>FlVFeFf+Fl
Ff<Fl
Vd Fe=kNe
V is velocity (regulating variable), /V dV dt= is acceleration, m is unknown vehicle mass,
eF is engine force, e eF kN= , where eN is torque, k is unknown conversion coefficient,
fF is friction force, fF V= ρ , where ρ is unknown friction coefficient,
lF is load force (unknown, dependent on the road profile).
The first order dynamics (Newton's Second Law):
( )e f l e lmV F F F kN V F= − + = − ρ − .
7
Define the state variable x V= , the control input eu N= , the disturbance /ld F m= − :
x ax bu d= − + + , (1)
y x v= + ,
where y is the output, v is the measurement noise, /a m= ρ , /b k m= .
Note: the engine from the introduction lecture has the same model .
Features:
− the constant parameters 0a > and 0b > are unknown ⇒ (1) is a variant of lΣ ;
− the time-varying signals d and v are unknown, but bounded;
− the unperturbed noise-free case: 0d v= = ,
− the reference signal dr V= , where dV − desired velocity.
Control problem (the asymptotic tracking):
( ) ( )x t r t→ with t → +∞ .
8
A variant of the solution:
1[ ]m mu b ay a y b r−= − + ,
where 0ma > and mb are parameters of the reference model:
m m m mx a x b r= − + .
The closed loop system has form:
m mx a x b r d= − + + , ( )md d a a v= + − .
In the noise-free case ( 0d v= = ⇒ 0d = ) the variable x has the desired dynamics!
To design the control u we have to estimate the unknown parameters a and b!
Let us try to solve this problem for the noise-free case. We will analyze the robustness issue
later. In this case the model (1) can be rewritten as follows:
y ay bu= − + . (1’)
9
2. ADAPTIVE PARAMETER ESTIMATION
a. Parameterized system model Consider a linear time-invariant SISO system described by the differential equation:
( )[ ]( ) ( )[ ]( )P s y t Z s u t= , (2)
( )y t ∈ , ( )u t ∈ are the measured output and input as before;
1
1 1 01
1 1 0
( ) ... ,( ) ... ,
n nn
m mm m
P s s p s p s pZ s z s z s z s z
−−
−−
= + + + += + + + +
are polynomials in s , with s being the differentiation operator
[ ]( ) ( )s x t x t= ;
ip , 0, 1i n= − and jz , 0,j m= are the unknown but constant parameters to be estimated.
Note: 1n = , 0m = ⇒ (1’) with 0p a= and 0z b= .
The objective: estimate the values ip , 0, 1i n= − and jz , 0,j m= using available for on-line measurements signals ( )y t and ( )u t (no a priori accessible datasets).
10
Parameterization:
let 11 1 0( ) ...n n
ns s s s−−Λ = + λ + + λ + λ be a stable polynomial (all zeros are in Re[ ] 0s < ).
Then (2) can be represented as follows:
( ) ( )[ ]( ) [ ]( )( ) ( )
P s Z sy t u ts s
=Λ Λ
⇒ ( ) ( ) ( )(1 )[ ]( ) [ ]( ) [ ]( )( ) ( ) ( )s P s Z sy t y t u ts s s
Λ− + =Λ Λ Λ
⇒
( ) ( ) ( )( ) [ ]( ) [ ]( )( ) ( )
Z s s P sy t u t y ts s
Λ −= +Λ Λ
. (3)
Define parameter vector * 1
0 1 1 0 0 1 1 2 2 1 1[ , ,..., , , , ,..., , ]T n mm m n n n nz z z z p p p p + +
− − − − −= λ − λ − λ − λ − ∈θ
and regressor function
1 11 1( ) [ ]( ),... [ ]( ), [ ]( ),... [ ]( )
( ) ( ) ( ) ( )
Tm n n ms st u t u t y t y ts s s s
− + +⎡ ⎤φ = ∈⎢ ⎥Λ Λ Λ Λ⎣ ⎦
.
Then (3) can be expressed in the equivalent form
*( ) ( )Ty t t= φθ . (4)
11
In (4): − the vector *θ contains all unknown parameters of the system (2);
− the regressor ( )tφ can be computed using the filters ( )
issΛ
, 0, 1i n= − .
Another variant of implementation:
1 12 2
( ) ( ) ( ),( ) ( ) ( ),t t u tt t y t
λλ
ω = ω +ω = ω +
A bA b
where 1( ) ntω ∈ , 2 ( ) ntω ∈ and
0 1 2 1
0 1 0 0 00 0 1 0 0
0 0 0 1n n
λ
− −
⎡ ⎤⎢ ⎥⎢ ⎥= ⎢ ⎥⎢ ⎥−λ −λ −λ −λ⎢ ⎥⎣ ⎦
A ,
0
01
⎡ ⎤⎢ ⎥
= ⎢ ⎥⎢ ⎥⎣ ⎦
b .
Then, we generate the regressor ( )tφ from
1 2( ) [ ( ) , ( ) ]T T Tmt t tφ = ω ωC ,
where ( 1)1 ( 1) ( 1)[ , ] m n
m m m n m+ ×
+ + × − −= ∈C I 0 ( 1 2( ) [ ( ) , ( ) ]T T Tt t tφ = ω ω for 1m n= − ).
12
b. Linear parametric model Linear parametric model has the form
*( ) ( )Ty t t= φθ , 0t t≥ , (4)
where * nθ∈θ is an unknown parameter vector, ( )y t ∈ is a known (measured) signal, ( ) nt θφ ∈ is a known vector signal (regressor), 1n n mθ = + + is the dimension of the model.
Features:
1) The model (4) is commonly seen in system modeling when unknown system parameters can be separated from known signals.
2) The components of ( )tφ may contain nonlinear and/or filtered functions of ( )y t and ( )u t (or some other system signals).
3) Adaptive parameter estimation based on ( )y t , ( )u t ⇔ Linear parametric model. Let ( )tθ be the estimate of *θ obtained from an adaptive update law, *( ) ( )t t= −θ θ θ is the parametric error, then define the estimation error
*( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )T T T Tt t t y t t t t t tε = φ − = φ − φ = φθ θ θ θ . (5)
13
Example 1
y ay bu= − + . (1’)
It has the form (2) for 0( )P s s p= + , 0( )Z s z= with 0p a= , 0z b= , 1m n= − , 1n = .
The filter 1 1( ) 1s s
=Λ +
.
The parameter vector * * * 21 2[ , ] [ ,1 ]T Tb a= θ θ = − ∈θ , 2nθ = .
The regressor 21 1( ) [ ]( ), [ ]( )1 1
Tt u t y t
s s⎡ ⎤φ = ∈⎢ ⎥+ +⎣ ⎦
.
The fast implementation 1 2( ) [ ( ), ( )]Tt t tφ = ω ω for
1 12 2
( ) ( ) ( ),( ) ( ) ( ),t t u tt t y t
ω = −ω +ω = −ω + 1λ = −A , 1=b .
The estimation error for the estimate 21 2( ) [ ( ), ( )]Tt t t= θ θ ∈θ :
1 1 2 2
1 1 2 2
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )( )( ( ) ) ( )( ( ) 1 ) ( ) ( ).
T
Tt t t y t t t t t y t
t t b t t a t tε = φ − = ω θ + ω θ − == ω θ − + ω θ − + = φ
θθ
14
c. Normalized gradient algorithm
How to update ( )tθ ? How to minimize the error ( ) ( ) ( ) ( ) ( ) ( )T Tt t t y t t tε = φ − = φθ θ ?
The idea is to choose the derivative of ( )tθ in a steepest descent direction in order to minimize
a normalized quadratic cost functional
2 * *
2 2 2( ) ( ) ( ) ( ) ( ) ( ( ) ) ( ) ( ) ( ( ) )( , )
2 ( ) 2 ( ) 2 ( )
T T T Tt t t t t t t t tJ tm t m t m tε φ φ − φ φ −= = =θ θ θ θ θ θθ ,
where ( )m t is a normalizing signal not depending (explicitly) on ( )tθ .
The idea of ( )m t choice: 2( ) ( ) / ( )Tt t m tφ φ has to be bounded (return later to this issue).
The steepest descent direction of ( , )J t θ is 2 2( , ) ( ) ( )( )
( ) ( )J t t tt
m t m t∂ ε φ∂ε− = − = −ε
∂ ∂θ
θ θ, therefore:
2( )( ) ( )( )tt t
m tφ= −εθ Γ , 0 0( )t =θ θ , 0t t≥ , (6)
where 0T= >Γ Γ is a design matrix gain, 0θ is an initial estimate of *θ .
15
For (6) an admissible choice of the normalizing function ( )m t is
( ) 1 ( ) ( )Tm t t t= + κφ φ ,
where 0κ > is a design parameter.
Example 1
The estimation error and the regressor:
1 1 2 2( ) ( ) ( ) ( ) ( ) ( )t t t t t y tε = ω θ + ω θ − , 1 2( ) [ ( ), ( )]Tt t tφ = ω ω .
The cost functional and derivative:
22
1 1 2 22 2
[ ( ) ( ) ( ) ( ) ( )]( )( , )2 ( ) 2 ( )
t t t t y ttJ tm t m t
ω θ + ω θ −ε= =θ , 12 2
( )( , ) ( )( )( )tJ t ttm t
ω∂ ε ⎡ ⎤= − ⎢ ⎥ω∂ ⎣ ⎦θ
θ.
The normalized gradient algorithm for 2= γΓ I , 0γ > and 1κ = :
12 2 21 2
( )( )( ) ( )1 ( ) ( )ttt tt t
ωε ⎡ ⎤= −γ ⎢ ⎥ω⎣ ⎦+ ω + ωθ .
16
Lemma 1. The adaptive algorithm (6) guarantees:
(i) ( )tθ , ( )tθ and ( ) / ( )t m tε are bounded (belong to L∞); (ii) ( ) / ( )t m tε and ( )tθ belong to 2L .
Proof. Introduce the positive definite (Lyapunov) function 1( ) TV −=θ θ Γ θ , then ( =θ θ)
2
1 12 2 2
( ) ( ) ( )2 2 ( ) 2 ( ) 2( ) ( ) ( )
TT T t t tV t t
m t m t m t− − ⎡ ⎤φ φ ε= = −ε = − ε = −⎢ ⎥
⎢ ⎥⎣ ⎦
θθ Γ θ θ Γ Γ , 0t t≥ . (7)
Since 0V ≤ we have: ( )V t L∞∈ ⇒ ( )t L∞∈θ ⇒ ( )t L∞∈θ = all these signals are bounded.
The boundedness of ( ) / ( )t m tε follows the boundedness of ( )tθ and the inequality
2
| ( ) | | ( ) ( ) | || ( ) || || ( ) ||( ) 1 ( ) ( ) 1 || ( ) ||
T
Tt t t t t
m t t t t
ε φ φ= ≤+ κφ φ + κ φ
θ θ .
Then boundedness of ( )tθ follows from the inequality
2 2( ) | ( ) | || ( ) || || ( ) || | ( ) ||| || ( ) || || || ||
( ) ( ) ( )( ) 1 || ( ) ||
t t t t ttm t m t m tm t t
φ ε φ φ ε= ε ≤ ≤+ κ φ
θ Γ Γ Γ . ⇒ (i)
17
Lemma 1. The adaptive algorithm (6) guarantees:
(i) ( )tθ , ( )tθ and ( ) / ( )t m tε are bounded (belong to L∞); (ii) ( ) / ( )t m tε and ( )tθ belong to 2L .
Proof. Let us rewrite the equality (7) in the form 2
2( )2 ( )( )t V t
m tε = − and integrate it:
0 0
21
0 0 0 02( )2 ( ) ( ) ( ) ( ) ( *) ( *)( )
t t Tt t
t dt V t dt V t V t V tm t
−ε− = = − ≤ = − − < ∞∫ ∫ θ θ Γ θ θ , 0t t≥ ,
therefore 2( )( )t L
m tε ∈ . From the inequality
2
|| ( ) || | ( ) ||| || || ||( )1 || ( ) ||
t tm tt
φ ε≤+ κ φ
θ Γ
we obtain that ( )tθ belongs to 2L . ⇒ (ii) ⇒ The Lemma 1 is proven.
Note: We did not prove that *lim ( )t t→∞ =θ θ !
18
Discussion:
1) The algorithm has equilibriums when || ( ) || 0t =θ , from (6) we have 2( )|| ( ) || || ( ) ||( )tt t
m tφ= εθ Γ :
|| ( ) || 0tφ = ⇒ || ( ) || 0t =θ ⇐ ( ) 0tε = ⇐ *( )t =θ θ ! *( )t =θ θ is not unique equilibrium of (6) (the usual drawback of any gradient algorithm)!
2) 1( ) ( ) ( )TV t t t−= θ Γ θ is a measure of deviation of ( )tθ from *θ , and from (7) ( ) 0V t ≤ ⇒ 1 1
0 0 0[ ( ) *] [ ( ) *] ( ) ( ) [ *] [ *]T Tt t V t V t− −− − = ≤ = − −θ θ Γ θ θ θ θ Γ θ θ .
3) From Lemma 1 we have that 2( ) / ( )t m t L L∞ε ∈ ∩ and lim ( ) / ( ) 0t t m t→∞ ε = .
4) From (7) we have that the function is nonincreasing ( ( ) 0V t ≤ ) and bounded from below ( ( ) 0V t ≥ ), thus there exists lim ( )t V t V→∞ ∞= for some constant 0V∞ ≥ : − 0V∞ = ⇒ *lim ( )t t→∞ =θ θ ;
− 0V∞ > ⇒ lim ( )t t→∞ ∞=θ θ for some constant vector nθ∞ ∈θ .
5) if ( )t L∞∈θ ⇒ 2( )t L L∞∈ ∩θ (Lemma 1) ⇒ lim ( ) 0t t→∞ =θ ⇒ lim ( )t t→∞ ∞=θ θ .
( ) sin( 1)t tθ = + , cos( 1)( ) 0.51
ttt
+θ =+
⇒ 2( )t L L∞θ ∈ ∩ , lim ( ) 0t t→∞ θ = , lim ( ) ?t t→∞ θ =
19
Example 1
Plant: y ay bu= − + .
Adaptive estimator: 1 1 2 2 12 2 21 2
( ) ( ) ( ) ( ) ( ) ( )( ) ( )1 ( ) ( )t t t t y t tt tt t
ω θ + ω θ − ω⎡ ⎤= −γ ⎢ ⎥ω⎣ ⎦+ ω + ωθ , 1 1
2 2
( ) ( ) ( ),( ) ( ) ( ).t t u tt t y t
ω = −ω +ω = −ω +
Simulation 1: 0.5a = , 1b = , 20γ = and ( ) sin( )u t t= ,
0 5 10 15 20
1−
0
1
2
y
u
t
0 5 10 15 200
0.5
1
1.5
1 a−
bθ1
θ2
t
Simulation 2: 1.5a = , 2b = , 20γ = and ( ) sin( )u t t= ,
0 5 10 15 20
2−
1−
0
1
2
y
u
t
0 5 10 15 201−
0
1
2
1 a−
b
θ1
θ2
t
20
Simulation 3: 1.5a = , 2b = , 20γ = and ( ) 1 cos( )tu t e t−= − ,
0 5 10 15 20
0
0.5
1
1.5
y
u
t
0 5 10 15 200
0.5
1
1.5
θ1
θ2
t
0 5 10 15 200.05
0.1
0.15
0.2
0.25
V
t
Conclusions:
− the convergence of adjusted estimates ( )tθ to their ideal values *θ depends on the input u ;
− y , u are oscillating ⇒ *( )t →θ θ ; y const→ , u const→ (set-point) ⇒ *( )t →θ θ .
21
d. Normalized lease-squares algorithm
2( ) ( )( ) ( )
( )t tt t
m tφ= −ε Pθ , 0 0( )t =θ θ , 0t t≥ , (8)
2( ) ( ) ( ) ( )( )
( )
Tt t t ttm t
φ φ= − P PP , 0 0 0( ) 0Tt = = >P P P , 0t t≥ , (9)
( ) 1 ( ) ( ) ( )Tm t t t t= + κφ φP ,
where 0κ > is a design parameter, 0θ is the initial estimate of *θ and 0P is the initial value of
the gain matrix ( ) n nt θ θ×∈P .
Note: − if ( )t =P Γ for all 0t t≥ , then (8) ⇒ (6); − the dimension of (6) is 1n n mθ = + + , as far as the dimension of (8), (9) is 2n nθ θ+ .
Example 1
1,1 1 1,2 22 2,1 1 2,2 2
P PP Pm
ω + ω⎡ ⎤ε= − ⎢ ⎥ω + ω⎣ ⎦θ , 1,1 1 1,2 2 1,1 1 1,2 2
2 2,1 1 2,2 2 2,1 1 2,2 2
1TP P P P
P P P Pm
ω + ω ω + ω⎡ ⎤ ⎡ ⎤= − ⎢ ⎥ ⎢ ⎥ω + ω ω + ω⎣ ⎦ ⎣ ⎦
P
2 2 21,1 1 1,2 1 1 2,1 1 1 2,2 21m P P P P= + ω + ω ω + ω ω + ω .
22
Lemma 2. The adaptive algorithm (8),(9) guarantees:
(i) ( ) ( ) 0Tt t= >P P for all 0t t≥ , ( )tP , ( )tP are bounded;
(ii) ( )tθ and ( ) / ( )t m tε are bounded (belong to L∞), where ( ) 1 ( ) ( )Tm t t t= + φ φ ; (iii) ( ) / ( )t m tε , ( ) / ( )t m tε and ( )tθ belong to 2L ;
(iv) there exist a constant matrix n nθ θ×∞ ∈P , and a constant vector nθ∞ ∈θ such that
lim ( )t t→∞ ∞=P P , lim ( )t t→∞ ∞=θ θ .
Proof. First, ( ) ( )Tt t=P P and ( )tP is bounded by the algorithm (9) construction:
( ) ( ) ( )( ) ( )1 ( ) ( ) ( )
T
Tt t tt t
t t tφ φ= −
+ κφ φPP P
P.
Second, the identity 1( ) ( ) nt t θ− =P P I implies
( )1 1 1 2( ) ( ) ( ) ( ) ( ) ( ) ( )Td t t t t m t t tdt
− − − −= − = φ φP P P P ,
then integrating this equality we obtain:
01 1 2
0( ) ( ) ( ) ( ) ( )t Ttt t m d− − −= + τ φ τ φ τ τ∫P P , 0t t≥ . (10)
10( ) 0t − >P ⇒ 1 1
0( ) ( ) 0t t− −≥ >P P ⇒ ( ) 0t >P and ( )tP is bounded. ⇒ (i)
23
Consider the positive definite function 1( , ) ( )TV t t −=θ θ P θ, then ( ( ) ( ) ( )Tt t tε = φθ )
( )1 1 1
1 12 2 2
2 2
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) (( ) ( )( ) ( )
T T T
T TT T
T T T
dV t t t t t t t t tdt
t t t t t tt t t t t t t tm t m t m tt t t t tt tm t m t
− − −
− −
= + + =
φ φ φ φ= −ε − ε + =
φ φ φ= −ε − ε +
θ P θ θ P θ θ P θ
P PP θ θ P θ θ
θ θ θ 202 2
) ( ) ( ) ( ) , .( ) ( )
Tt t t t t tm t m t
φ ε= − ≥θ
(11)
Hence, ( ) [ , ( )]V t V t t= θ is bounded, and using (10) we obtain:
( )01 2
0( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )tT T TtV t t t t t m d t− −= + τ φ τ φ τ τ < ∞∫θ P θ θ θ , 0t t≥ .
Therefore 1
0( ) ( ) ( )Tt t t−θ P θ is bounded ⇒ ( )tθ and ( )tθ are bounded.
Boundedness of ( ) / ( )t m tε follows the proven property ( )t L∞∈θ and the inequality
| ( ) | | ( ) ( ) | || ( ) || || ( ) ||( ) 1 ( ) ( ) 1 ( ) ( )
T
T Tt t t t t
m t t t t t
ε φ φ= ≤+ φ φ + φ φ
θ θ . ⇒ (ii)
24
Rewriting the equality (11) in the form 2 22 ( ) / ( ) ( )t m t V tε = − and integrating it, we obtain:
0 0
21
0 0 0 0 02( )2 ( ) ( ) ( ) ( ) ( *) ( *)( )
t t Tt t
t dt V t dt V t V t V tm t
−ε− = = − ≤ = − − < ∞∫ ∫ θ θ P θ θ , 0t t≥ ,
therefore 2( )( )t L
m tε ∈ and
( ) ( ) ( )( ) ( ) ( )t t m t
m t m t m tε ε= + 2
( )( )t L
m tε ∈ , ( )
( )m t Lm t ∞∈ ⇒ 2
( )( )t L
m tε ∈ .
Since ( ) ( )Tt t=P P is bounded and ( ) ( ) ( )s st t t=P P P ( ( )s tP is also bounded) we have
2
2
( ) ( ) ( ) ( ) | ( ) ||| ( ) || ( )( )( ) 1 ( ) ( ) ( )
( ) ( ) ( ) || ( ) ( ) ||| ( ) | | ( ) ||| ( ) || ,( ) ( )1 ( ) ( ) ( ) ( ) 1 || ( ) ( ) ||
T
s s ssT
s s s
t t t t tt tm tm t t t t
t t t t tt ttm t m tt t t t t t
φ φ ε= ε = =+ κφ φ
φ φε ε= =+ κφ φ + κ φ
P PθP
P P PPP P P
therefore, 2( )t L∈θ . ⇒ (iii)
25
The integration of the differential equation (9) gives for 0t t≥ :
00 2
( ) ( ) ( ) ( )( ) ( ) 0( )
Tttt t d
mτ φ τ φ τ τ= − τ >
τ∫
P PP P ⇒ 00 2
( ) ( ) ( ) ( )( )( )
Tttt d
mτ φ τ φ τ τ> τ
τ∫
P PP .
For any nθ∈z we have 00 2
( ) ( ) ( ) ( )( ) 0( )
TtT Ttt d
mτ φ τ φ τ τ∞ > > τ ≥
τ∫
P Pz P z z z , consequently, the
scalar function 0 2
( ) ( ) ( ) ( )( , )( )
Tt Ttf t d
mτ φ τ φ τ τ= τ
τ∫
P Pz z z has properties:
− it is a nondecreasing function of 0t t≥ ; − it is upper and lower bounded,
then there exists f ∈z such that lim ( , )t f t f→∞ = zz . ⇒ lim ( )t t→∞ ∞=P P , n nθ θ×∞ ∈P .
Note that ( ( ) ( ) ( )Tt t tε = φ θ )
1 12 2
T
m m− −φ φ= = − ε = − φ =P Pθ θ PP θ PP θ ⇒ 1
0 0( ) ( ) ( ) ( )t t t t−=θ P P θ ⇒
* 1 * 10 0 0 0lim ( ) lim ( ) ( ) ( ) ( ) ( ) n
t tt t t t t t θ− −→∞ →∞ ∞ ∞= + = + = ∈θ θ P P θ θ P P θ θ .
26
Discussion:
1) The algorithm (8)−(9) can be presented in the form 1
0( ) ( ) ( ) ( ) ( , ) ( )t t t t t t t−= =θ P P θ A θ , 10 0( ) ( ) ( ) ( )t t t t−=θ P P θ ,
thus it is a linear time-varying system!!! The same as the algorithm (6):
2 2( ) ( ) ( )( ) ( ) ( ) ( ) ( ) ( )( ) ( )
Tt t tt t t t t tm t m tφ φ φ= = −ε = − =θ θ Γ Γ θ B θ .
2) Uniform stability: 10 0 0 0|| ( ) || || ( ) ( ) ( ) || || ( ) ||t t t t c t−= ≤θ P P θ θ for some 0 0c > .
3) The least-squares algorithm (8), (9) minimizes a cost function which is an integral of squared errors at many time instants with a penalty on the initial estimate 0 0( )t =θ θ :
0
0
21
0 0 02
21
0 0 02
( ( ) ( ) ( ))1 1( , ) [ ( ) ] [ ( ) ]2 2( )
( )1 1 ( ) ( ).2 2( )
Tt Tt
t Tt
yJ t d t tm
d t tm
−
−
τ φ τ − τ= τ + − − =τ
ε τ= τ +τ
∫
∫
θθ θ θ P θ θ
θ P θ
Compare with the gradient descent algorithm (6): 2
2( )1( , )
2 ( )tJ t
m tε=θ .
27
Example 1 Plant: y ay bu= − + .
Estimator: 1,1 1 1,2 22 2,1 1 2,2 2
P PP Pm
ω + ω⎡ ⎤ε= − ⎢ ⎥ω + ω⎣ ⎦θ , 1,1 1 1,2 2 1,1 1 1,2 2
2 2,1 1 2,2 2 2,1 1 2,2 2
1TP P P P
P P P Pm
ω + ω ω + ω⎡ ⎤ ⎡ ⎤= − ⎢ ⎥ ⎢ ⎥ω + ω ω + ω⎣ ⎦ ⎣ ⎦
P
Simulation 1: 0.5a = , 1b = , 0 220=P I and ( ) sin( )u t t= ,
0 5 10 15 20
1−
0
1
2
y
u
t
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1 a−
b
θ1
θ2
t
Simulation 2: 1.5a = , 2b = , 0 250=P I and ( ) sin( )u t t= ,
0 5 10 15 20
2−
1−
0
1
2
y
u
t
0 5 10 15 201−
0
1
2
1 a−
b
θ1
θ2
t
28
Simulation 3: 1.5a = , 2b = , 0 250=P I and ( ) 1 cos( )tu t e t−= − , ( ) || ( ) ||t tσ = P ,
0 5 10 15 200
0.5
1
1.5
y
u
t
0 5 10 15 200
0.2
0.4
0.6
0.8
1
θ1
θ2
t
0 5 10 15 200.03
0.04
0.05
0.06
0.07
0.08
0.09
V
t
0 5 10 15 2035
40
45
50
55
σ
t
Conclusions:
− the rate of convergence in the algorithm (8), (9) is a more complex issue than in (6);
− the convergence of adjusted estimates ( )tθ to their ideal values *θ depends on the input u ;
− y , u are oscillating ⇒ *( )t →θ θ ; y const→ , u const→ (set-point) ⇒ *( )t →θ θ .
29
e. Discrete-time version of adaptive algorithms
Continuous time 0t t≥ ⇒ Discrete time 0 0 0 , , 2 ...t t t T t T∈ + + , 0T > is the period.
The normalized gradient algorithm:
2( ) ( )( 1) ( )
( )t tt t
m tφ ε+ = −θ θ Γ , 0 0( )t =θ θ , 2 0T
nθ > = >I Γ Γ ,
( ) ( ) ( )Tm t t t= κ + φ φ , 0κ > .
The normalized least-squares algorithm:
2( ) ( )( 1) ( ) ( 1)
( )t tt t t
m tφ ε+ = − −θ θ P , 0 0( )t =θ θ ,
2( ) ( )( ) ( 1) ( 1) ( 1)
( )
Tt tt t t tm t
φ φ= − − − −P P P P , 0 0 0( 1) 0Tt − = = >P P P ,
( ) ( ) ( 1) ( )Tm t t t t= κ + φ − φP , 0κ > . Proprieties:
− ( )tθ , ( ) / ( )t m tε , ( ) / ( )t m tε and ( ) ( ) 0Tt t= >P P are bounded; − ( ) / ( )t m tε , ( ) / ( )t m tε and ( 1) ( )t t+ −θ θ belong to 2L .
30
3. IDENTIFICATION AND ROBUSTNESS
− identification ⇔ parameter convergence; − robustness ⇔ || || 0≠d , || || 0≠v .
a. Parametric convergence and persistency of excitation Lemma 3. For the gradient algorithm (6) or least-squares algorithm (8)−(9), if ( )m t L∞∈
and ( )t L∞φ ∈ , then lim ( ) 0t t→∞ ε = .
Proof. 2( )( )t L L
m t ∞ε ∈ ∩ and ( ), ( )t t L∞∈θ θ from lemmas 1, 2 . Since ( ) ( ) ( )Tt t tε = φ θ we have
( ) ( ) ( ) ( ) ( )T Tt t t t tε = φ + φθ θ . Hence: ( )t L∞φ ∈ ⇒ ( )t L∞ε ∈ , ( )m t L∞∈ ⇒ 2( )t L L∞ε ∈ ∩ .
Under conditions of lemma 3 asymptotically *( ) [ ( ) ] ( ) 0ni i iit t tθε = θ − θ φ =∑ , 1t t≥ :
a) ( ) [1,0,...,0]Ttφ = ⇒ *1 1( ) 0tθ − θ = , ( )i tθ for 2 i nθ≤ ≤ −?
b) ( ) [1,1,...,1]Ttφ = ⇒ *[ ( ) ] 0ni ii tθ θ − θ =∑ −?
c) ( ) sin( )i t itφ = ω , 1,i nθ= , 0ω > ⇒ *[ ( ) ]sin( ) 0ni ii t itθ θ − θ ω =∑ ⇒ *( )i itθ = θ , 1,i nθ= .
31
Definition 1. A bounded vector signal ( ) qtϕ ∈ , 1q ≥ , is exciting over the finite time interval
0 0 0[ , ]σ σ + δ , 0 0δ > , 0 0tσ ≥ , if for some 0 0α >
0 00 0( ) ( )T qdσ +δ
σ ϕ τ ϕ τ τ ≥ α∫ I .
Definition 2. A bounded vector signal ( ) qtϕ ∈ , 1q ≥ , is Persistently Exciting (PE) if there
exist 0δ > and 0α > such that
( ) ( )T qdσ+δσ ϕ τ ϕ τ τ ≥ α∫ I , 0t∀σ ≥ .
( ) qtϕ ∈ is PE ⇔ 0∃ρ > , 0δ > : 0 0( ) ( ) ( )t T
qt d t tϕ τ ϕ τ τ ≥ ρ −∫ I , 0t t∀ ≥ + δ
(positive definite in average).
The idea: rank[ ( ) ( ) ] 1Tt tϕ ϕ = , 0t t≥ ⇒ 00
rank[ ( ) ( ) ]t Tt d q+δϕ τ ϕ τ τ =∫ .
Example 2.
( ) [1,1]Ttϕ = ⇒ 1 1( ) ( ) 1 1Tt t ⎡ ⎤ϕ ϕ = ⎢ ⎥⎣ ⎦
⇒ 01 1( ) ( ) 01 1
T dδ ⎡ ⎤ϕ τ ϕ τ τ = δ ≥⎢ ⎥⎣ ⎦∫ ⇒ not PE.
32
( ) [1, ]t Tt e−ϕ = ⇒ 0 2e( ) ( )
e 0.5eT d
−δδ−δ − δ
⎡ ⎤δ −ϕ τ ϕ τ τ = ⎢ ⎥− −⎢ ⎥⎣ ⎦
∫ ⇒ exciting over some finite intervals.
( ) [1,sin( )]Tt tϕ = ⇒ 20cos( )( ) ( ) ( )cos( ) 0.5 0.25sin(2 )
T dδ δ − δ⎡ ⎤ϕ τ ϕ τ τ = ≥ λ δ⎢ ⎥− δ − δ − δ⎣ ⎦∫ I ,
2
26 sin(2 ) [2 sin(2 )]( ) cos( )8 64
δ − δ δ + δλ δ = − + δ ≥ ρδ , 0.4ρ = for 5δ > :
0 2 4 6 8 10
2−
2
4
6
λ t( )
ρ t⋅
t
cos( )( ) sin( )tt t
⎡ ⎤ϕ = ⎢ ⎥⎣ ⎦ ⇒
220 2
0.5sin(2 ) sin( )1 1( ) ( ) [ sin( )]2 2sin( ) 0.5sin(2 )
T dδ ⎡ ⎤δ + δ δϕ τ ϕ τ τ = ≥ δ − δ⎢ ⎥δ δ − δ⎢ ⎥⎣ ⎦
∫ I ⇒ PE.
⇒ PE!!!
33
Normalized gradient algorithm (6) ( *( ) ( )t t= −θ θ θ , ( ) ( ) ( )Tt t tε = φ θ ):
2 2( ) ( )( ) ( ) ( ) ( ) ( ) ( ) ( )( ) ( )
Tt tt t t t t t tm t m tφ φ= = −ε = − φ =θ θ Γ Γ θ B θ , 2
( ) ( )( )( )
Tt ttm t
φ φ= −B Γ .
Let 0( , )t tΦ be the state transition matrix of the linear time-varying system (6), then
− 0 0( ) ( , ) ( )t t t t= Φθ θ ;
− ( )tφ is PE ⇒ ( ) / ( )t m tφ , ( ) 1 ( ) ( )Tm t t t= + κφ φ is PE ⇒ 0( ) ( , ) ( ) / ( )Tt t t t m tη = Φ φ is PE:
0∃ρ > , 0δ > : 0 0( ) ( ) ( )t T
nt d t t θη τ η τ τ ≥ ρ −∫ I , 0t t∀ ≥ + δ.
Consider the Lyapunov function 1( ) TV −=θ θ Γ θ :
2
0 02 2( ) ( ) ( )2 2 ( ) ( ) 2 ( ) ( )( ) ( )
TT T Tt t tV t t t t
m t m tε φ φ= − = − = − η ηθ θ θ θ ,
integrating this equality for 0t t≥ + δ we obtain ( 10 0 0( ) TV t −= θ Γ θ ):
01
0 0 0 0 0 0 0 0 0 0( ) ( ) 2 ( ) ( ) ( ) 2 ( ) [ 2 ( )]tT T T TtV t V t d V t t t t t−= − η τ η τ τ ≤ − ρ − = − ρ −∫θ θ θ θ θ Γ θ ⇒
lim ( ) 0t V t→∞ = ⇒ *lim ( )t t→∞ =θ θ .
34
Normalized least-squares algorithm (8)−(9):
10 0( ) ( ) ( ) ( )t t t t−=θ P P θ , 0t t≥ .
Properties:
− lim ( ) 0t t→∞ =P ⇔ lim ( ) 0t t→∞ =θ ;
− 00 2
( ) ( ) ( ) ( )( ) ( )( )
Tttt t d
mτ φ τ φ τ τ= − τ
τ∫
P PP P , ( ) ( ) 0Tt t= >P P for all 0t t≥ ;
− ( )tφ is PE ⇒ ( ) / ( )t m tφ , ( ) 1 ( ) ( )Tm t t t= + κφ φ is PE ⇒ ( ) ( ) ( ) / ( )t t t m tη = φP is PE:
0∃ρ > , 0δ > : 0 0( ) ( ) ( )t T
nt d t t θη τ η τ τ ≥ ρ −∫ I , 0t t∀ ≥ + δ.
Then
00 0 00 ( ) ( ) ( ) ( ) ( ) ( ) 0t Tntt t d t t t θ< = − η τ η τ τ ≤ − ρ − ≤∫P P P I for some 0t t≥ ⇒
*lim ( )t t→∞ =θ θ .
Lemma 4. For the gradient algorithm (6) or least-squares algorithm (8)−(9), if ( )tφ is PE, then *lim ( )t t→∞ =θ θ .
35
Discussion:
What is PE property of the regressor ( )tφ :
1 2( ) [ ( ) , ( ) ]T T Tmt t tφ = ω ωC ,
where 1( ) ntω ∈ , 2 ( ) ntω ∈ and for a Hurwitz matrix λA :
1 12 2
( ) ( ) ( ),( ) ( ) ( ).t t u tt t y t
λλ
ω = ω +ω = ω +
A bA b
PE of ( )tφ ⇐ PE of 1( )tω and 2( )tω ⇐ PE of ( )u t and ( )y t .
(2) is a linear system ⇒ PE of ( )y t is determined by the input ( )u t !
PE of ( )u t ⇒ PE of ( )tφ
(that we already observed in the example).
36
Example 1 Plant: y ay bu= − + , 0.5a = , 1b = and ( ) sin( )u t t= .
Gradient algorithm: 20γ =
0 5 10 15 20
1−
0
1
2
y
u
t
0 5 10 15 200
0.5
1
1.5
1 a−
bθ1
θ2
t
Least-squares algorithm: 0 220=P I
0 5 10 15 20
1−
0
1
2
y
u
t
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1 a−
b
θ1
θ2
t
( ) sin( )u t t= ⇒ ( ) sin( )y t t= α + β ⇒ ( ) sin( )i i it tω = α + β due to 1 12 2
( ) ( ) ( ),( ) ( ) ( )t t u tt t y t
ω = −ω +ω = −ω + ⇒
1 2( ) [ ( ), ( )]Tt t tφ = ω ω ⇒ ( ) [cos( ),sin( )]Tt t tϕ = ⇒ PE.
37
b. Robustness of adaptive algorithms Before the noise free case with ( ) 0t =d and ( ) 0t =v has been considered for
( ) ( ) ;: .l= + +Σ = + +
x A θ x B θ u dy Cx Du v
What happens if ( ) 0t ≠d or ( ) 0t ≠v ? (only the case ( ) 0t ≠d will be considered)
Example 1 Plant: ( )y ay bu d t= − + + ,
1.5a = , 2b = and ( ) sin( )u t t= , ( ) 0.5sin(3 )d t t= .
0 100 2002−
1−
0
1
2
y
u
t
0 100 2001−
0
1
2
3
1 a−
bθ1
θ2
t
0 100 2001 10 3−×
0.01
0.1
1
V
t
0 100 2002−
1−
0
1
2
y
u
t
0 100 2001−
0
1
2
1 a−
b
θ1
θ2
t
0 100 2000
0.02
0.04
0.06
0.08
0.1
V
t
0 100 2000
10
20
30
40
50
σ
t
( )tφ is PE ⇒ Robustness!!!
(6) (8)−(9)
38
( ) 1 cos( )tu t e t−= −
0 100 2000
0.5
1
1.5
2
y
u
t
0 100 2000
0.5
1
1.5
θ1
θ2
t
0 100 2000.01
0.1
1
V
t
0 100 2000
0.5
1
1.5
2
y
u
t
0 100 2000
0.2
0.4
0.6
0.8
1
θ1
θ2
t
0 100 2000.02
0.04
0.06
0.08
0.1
0.12
V
t
0 100 2000
10
20
30
40
50
σ
t
( ) sin( )u t t= , ( ) 0.5sin( )d t t=
0 100 2002−
1−
0
1
2
y
u
t
0 100 2001−
0
1
2
3
1 a−
bθ1
θ2
t
0 100 2001 10 3−×
0.01
0.1
1
V
t
0 100 2002−
1−
0
1
2
y
u
t
0 100 2001−
0
1
2
3
1 a−
bθ1
θ2
t
0 100 2000
0.02
0.04
0.06
0.08
0.1
V
t
0 100 2000
10
20
30
40
50
σ
t Conclusion: the disturbance can seriously modify the system behavior.
(6) (8)−(9)
(6) (8)−(9)
39
Linear parametric model with modeling errors:
*( ) ( ) ( )Ty t t t= φ + δθ , 0t t≥ ,
where * nθ∈θ is an unknown parameter vector, ( ) nt θφ ∈ is a known regressor, ( )y t ∈
is a measured output, ( )tδ ∈ represents system modeling errors:
1 2| ( ) | || ( ) ||t c t cδ ≤ φ + , 1 0c > , 2 0c > .
Let ( ) nt θ∈θ be the estimate of *θ and define the estimation error
( ) ( ) ( ) ( ) ( ) ( ) ( )T Tt t t y t t t tε = φ − = φ + δθ θ , 0t t≥ ,
where *( ) ( )t t= −θ θ θ is the parametric error.
Modified gradient algorithm (6):
2( )( ) ( ) ( )( )tt t t
m tφ= −ε +θ Γ Γf , 0 0( )t =θ θ , ( ) 1 ( ) ( )Tm t t t= + κφ φ , 0κ > , 0t t≥ , (12)
where 0T= >Γ Γ is a design matrix gain, ( ) nt θ∈f is the modification term for robustness.
40
Stability & robustness analysis for nonlinear systems ⇔ Lyapunov function theory
11( )2
TV −=θ θ Γ θ , 2
2 2( ) ( ) ( ) ( )( ) ( )
Tt t tV tm t m tε ε δ= − + + θ f
Note:
1 2 1 2 12
|| ( ) ||| ( ) |( ) ( )1 ( ) ( )T
c t c c c ct cm t m tt t
φ +δ ≤ ≤ + ≤ +κ κ+ κφ φ
.
Then
2
1 22
( ) | ( ) | ( )( ) ( )( )
Tc ct tV tm t m tm t
⎡ ⎤ε ε≤ − + + +⎢ ⎥κ⎣ ⎦θ f ,
and
1 2| ( ) |( ) ( )
c ctm t m tε ≥ +
κ ⇒
21 2
2( ) | ( ) |
( ) ( )( )c ct t
m t m tm t
⎡ ⎤ε ε− ≤ − +⎢ ⎥κ⎣ ⎦ ⇒ ( )TV t≤ θ f .
The simplest modification:
2( )( ) ( )( )
stt f t
m tφ=f , 1 20 if | ( ) | / ( ) / / ( ),( )
( ) otherwise.st m t c c m tf t
t⎧ ε ≥ κ += ⎨ε⎩
⇒ 0V ≤ .
41
The simplest modification:
2( )( ) ( )( )
stt f t
m tφ=f , 1 2( ) if | ( ) | / ( ) / / ( ),( )
0 otherwise.st t m t c c m tf t ⎧ε ε < κ += ⎨
⎩ ⇒ 0V ≤ .
A dead zone modification:
2( )( ) ( )( )
dtt f t
m tφ=f , 1 2
1 2
( ) if | ( ) | / ( ) / / ( ),( )[ ( ) / ] [ ( )] otherwise.d
t t m t c c m tf tc m t c sign t
⎧ε ε < κ += ⎨κ + ε⎩
⇒ 0V ≤ .
2( )( ) [ ( ) ( )]( )
dtt t f t
m tφ= − ε −θ Γ ⇒
σ-Modification:
( ) ( )t t= −σf θ ⇒ 2( )( ) ( ) ( )( )tt t t
m tφ= −σ − εθ Γθ Γ ⇒ ( )t L∞∈θ .
2− 1− 0 1 2
2−
1−
1
2
ε
ε fs ε( )−
ε fd ε( )−
ε
42
Projection: assume that the set of admissible values for *θ is given, i.e.
* :|| || n Mθ∈ Ω = ∈ ≤θ θ θ , 0M > .
Projection has to ensure that ( )t ∈ Ωθ for all 0t t≥ , therefore
2
2
( )0 if || ( ) || or || ( ) || and ( ) ( ) 0,( )
( )( ) ( ) ( ) ( ) otherwise.
( ) ( ) ( )
T
T
T
tt M t M t tm t
tt t t t
t t m t
φ⎧ < = ε ≤⎪⎪= ⎨ φ⎪ ε⎪⎩
θ θ θ Γ
fΓθ θ Γθ Γθ
2θ
1θ
*θ
|| || M=θ
Inside the circle doing nothing.
On an attempt to exit the circle.
43
The properties: − boundedness of ( )tθ , ( )tθ and ( ) / ( )t m tε (belong to L∞); − ( ) / ( )t m tε and ( )tθ belong to 2L ; − in the noise-free case ( ( ) 0t =d ) the quality is preserved? ⇒ ESTIMATION?
Example 1 Plant: ( )y ay bu d t= − + + ,
0.5a = , 1b = and ( ) ( ) 1 cos( )tu t u t e t−= = − ; (6) with 20γ = .
0 100 2000
0.5
1
1.5
2
y
u
t
0 100 2000
0.5
1
1.5
1 a−
bθ1
θ2
t
( ) 0.5sin(0.3 )d t t=
0 100 2000
1
2
3
y
u
t
0 100 2001−
0
1
2
1 a−
bθ1
θ2
t
ROBUSTNESS!
44
Dead zone algorithm:
1− 0 1
1−
1θ2
1 a−
n x( )
n x( )−
θ1 b, x,
0 100 2000
0.5
1
1.5
1 a−
bθ1
θ2
t
σ-Modification (σ = 0.01):
1− 0 1
1−
1θ2
1 a−
n x( )
n x( )−
θ1 b, x,
0 100 2000
0.5
1
1.5
1 a−
bθ1
θ2
t
Projection (M = 1.5):
1− 0 1
1−
1θ2
1 a−
n x( )
n x( )−
θ1 b, x,
0 100 2000.5−
0
0.5
1
1.5
2
1 a−
bθ1
θ2
t
d(t)=
0.5
sin(
0.3t
)
θ1
θ2
θ1
θ2
θ1
θ2
45
Dead zone algorithm:
1− 0 1
1−
1θ2
1 a−
n x( )
n x( )−
θ1 b, x,
0 100 2000
0.2
0.4
0.6
0.8
1
1 a−
b
θ1
θ2
t
σ-Modification (σ = 0.01):
1− 0 1
1−
1θ2
1 a−
n x( )
n x( )−
θ1 b, x,
0 100 2000
0.5
1
1.5
1 a−
bθ1
θ2
t
Projection (M = 1.5):
1− 0 1
1−
1θ2
1 a−
n x( )
n x( )−
θ1 b, x,
0 100 2000
0.5
1
1.5
1 a−
bθ1
θ2
t
d(t)=
0
θ1
θ2
θ1
θ2
θ1
θ2
46
SUMMARY
1. Adaptive parameter estimation:
a. Parameterized system model *( ) ( )Ty t t= φ θ .
b. Linear parametric model ( ) ( ) ( ) ( ) ( ) ( )T Tt t t y t t tε = φ − = φθ θ , *( ) ( )t t= −θ θ θ .
c. Normalized gradient algorithm 2( )( ) ( )( )tt t
m tφ= −εθ Γ .
d. Normalized least-squares algorithm 2( ) ( )( ) ( )
( )t tt t
m tφ= −ε Pθ , 2
( ) ( ) ( ) ( )( )( )
Tt t t ttm t
φ φ= − P PP .
e. Discrete-time version of adaptive algorithms.
2. Identification and robustness:
f. Parametric convergence and PE (PE ⇒ convergence/estimation ⇒ robustness).
g. Robustness of adaptive algorithms (robustness ⇔ estimation).
47
Example 2
Oscillating pendulum:
ϕ f, d
[ , )ϕ∈ −π π is the pendulum angle, f ∈ is the (controlling or exciting) input applied to the
support, d ∈ is the disturbance influencing the support also.
Nonlinear model:
2 sin( ) cos( ) ( ) ( )y y y b y f t d t= −ω − ρ + + , (13)
[ , )y = ϕ∈ −π π is the measured angle, y ∈ and y ∈ are the angle velocity and acceleration; 0ρ > is an unknown friction coefficient, 0ω > is an unknown natural frequency,
0b > is an unknown control gain.
3 unknown parameters + nonlinearity. ⇒ Define 1 sin( )u y= and 2 cos( )u y u= :
21 2( ) ( ) ( )y y u t bu t d t+ ρ = −ω + + ⇒ (2) for 2n = , 1m = and a vector 1 2[ , ]Tu u=u .
48
Define the polynomials: 2
1( )P s s p s= + , 1p = ρ ; 21 1,0( )Z s z= = −ω ; 2 2,0( )Z s z b= = ,
then the noise-free model (13) has the form 1 1 2 2( )[ ]( ) ( )[ ]( ) ( )[ ]( )P s y t Z s u t Z s u t= + .
Parameterization for 21 0( )s s sΛ = + λ + λ :
1 21 2
( ) ( )( ) [ ]( ) [ ]( ) [ ]( )( ) ( ) ( )
Z s Z sP s y t u t u ts s s
= +Λ Λ Λ
⇒
1 21 2
( ) ( )( ) ( )(1 )[ ]( ) [ ]( ) [ ]( ) [ ]( )( ) ( ) ( ) ( )
Z s Z ss P sy t y t u t u ts s s s
Λ− + = +Λ Λ Λ Λ
⇒
1 21 2
( ) ( )( ) ( )( ) [ ]( ) [ ]( ) [ ]( )( ) ( ) ( )
Z s Z ss P sy t y t u t u ts s s
Λ −= + +Λ Λ Λ
⇒ *( ) ( )Ty t t= φ θ ,
the parameterized system model for 10( ) ( ) ( )[ ]( )y t y t s y t−= − λ Λ , * 2
1[ , , ]Tb= λ − ρ −ωθ and
1 2 0,2 1,1 2,11 1( ) [ ]( ), [ ]( ), [ ]( ) [ , , ]( ) ( ) ( )T Tst y t u t u ts s s
⎡ ⎤φ = = ω ω ωΛ Λ Λ⎢ ⎥⎣ ⎦, 0 0,1( ) ( ) ( )y t y t t= − λ ω ,
0 01 1 12 2 2
( ) ( ) ( ),( ) ( ) ( ),( ) ( ) ( ),
t t y tt t u tt t u t
λλλ
ω = ω +ω = ω +ω = ω +
A bA bA b
0 1
0 1λ
⎡ ⎤= ⎢ ⎥−λ −λ⎣ ⎦A , 0
1⎡ ⎤= ⎢ ⎥⎣ ⎦
b .
49
1ω = , 0.1ρ = , 0.5b = , ( ) sin(3 )f t t= , 0 1λ = , 1 2λ = , 100γ = .
( ) 0d t = ( ) 0.5sin(0.3 )d t t=
0.3− 0.2− 0.1− 0 0.1 0.2 0.30.4−
0.2−
0
0.2
0.4
ϕ'
ϕ
1− 0.5− 0 0.5 10.4−
0.2−
0
0.2
0.4
ϕ'
ϕ
0 20 40 60 80 1002−
1−
0
1
2
ω2−
λ1 ρ−
θ1
θ2
θ3
b
t
0 20 40 60 80 1002−
1−
0
1
2
3
ω2−
λ1 ρ−θ1
θ2
θ3
b
t
0 20 40 60 80 1002−
1−
0
1
2
ω2−
λ1 ρ−
θ1
θ2
θ3
b
t
0 20 40 60 80 1002−
1−
0
1
2
3
ω2−
λ1 ρ−θ1
θ2
θ3
b
t
Nor
mal
ized
gra
dien
t al
gori
thm
D
ead
zone
m
odifi
catio
n
50
4. INDIRECT ADAPTIVE CONTROL
Adjustment of control parameters:
− direct (from an adaptive control law/Lyapunov analysis); − indirect (from adaptive estimates of the system parameters).
Indirect adaptive control design:
1) adaptive estimation of the plant parameters; 2) calculation of control parameters.
a. Model reference control The main steps:
1) adaptive estimation algorithm design;
2) reference model selection;
3) controller structure construction;
4) controller parameter calculation;
5) stability and robustness analysis.
Pm(θ* )
u y C(θс)
Adaptive parameter estimation
Control parameter derivation
( )tθ
θc(t)
51
Example 1
Plant: y ay bu d= − + + .
Adaptive estimation algorithm ( * * *1 2[ , ] [ ,1 ]T Tb a= θ θ = −θ ):
12 2
( )( )( ) ( )( )ttt tm t
ωε ⎡ ⎤= −γ ⎢ ⎥ω⎣ ⎦θ , 2 2
1 2( ) 1 ( ) ( )m t t t= + ω + ω , 1 12 2
( ) ( ) ( ),( ) ( ) ( ),t t u tt t y t
ω = −ω +ω = −ω + 1λ = −A , 1=b .
Reference model: ( )m m m my a y b r t= − + where ( )r t ∈ is the reference signal to be tracked, 0ma > (the reference model is stable).
Controller structure: 1[( ) ]m mu b a a y b r−= − + ⇒ m my a y b r d= − + + .
Controller parameter calculation:
1 2c cu y r= θ + θ , 1
1 1 2(1 )cma−θ = θ − θ − , 1
2 1c
mb−θ = θ .
Division on 1θ ⇒ projection modification of the adaptation algorithm:
1 12 2
( ) ( )( )( ) ( ) 0( )t f ttt tm t
ωε ⎡ ⎤ ⎡ ⎤= −γ +⎢ ⎥ ⎢ ⎥ω ⎣ ⎦⎣ ⎦θ , 1 min 1 min 1
1 21
0 if ( ) or ( ) and ( ) ( ) 0,( )
( ) ( ) ( ) otherwise.t b t b t t
f tt t m t −
θ > θ = ε ω ≥⎧= ⎨γ ε ω⎩
min 0b > is the low bound for b , i.e. minb b≥ .
52
1.5a = , 2b = , 1ma = , 1mb = , min 0.1b =
0 5 10 15 20
1−
1y
u
r
t
0 5 10 15 200.5−
0
0.5
1
1.5
1 a−
θ1
θ2
t
0 5 10 15 20
1−
1y
u
r
t
0 5 10 15 201−
0
1
2
1 a−
b
θ1
θ2
t
0 5 10 15 20
1−
1y
u
r
t
0 5 10 15 200.5−
0
0.5
1
1.5
1 a−
θ1
θ2
t
( ) 0d t =
( ) 0.5sin(3 )d t t=
( ) 0.5sin(0.3 )d t t=
53
The general procedure: ( )[ ]( ) ( )[ ]( ) ( )pP s y t k Z s u t d t= + , 0t ≥ , (14)
( )y t ∈ , ( )u t ∈ are the measured output and input as before;
11 1 0( ) ...n n
nP s s p s p s p−−= + + + + , 1
1 1 0( ) ...m mmZ s s z s z s z−
−= + + + + ,
pk , ip , 0, 1i n= − and jz , 0, 1j m= − are the unknown but constant parameters.
Assumption 1. The constant min| | 0pk k≥ > and ( )psign k are given. ⇒ Necessary.
Assumption 2. pk k k≤ ≤ ; i i ip p p≤ ≤ , 0, 1i n= − ; j j jz z z≤ ≤ , 0, 1j m= − . ⇒ Desired.
1) Adaptive estimation algorithm design:
( ) ( ) ( )( ) [ ]( ) [ ]( )( ) ( )p
Z s s P sy t k u t y ts s
Λ −= +Λ Λ
, 11 1 0( ) ...n n
ns s s s−−Λ = + λ + + λ + λ ⇒
*( ) ( )Ty t t= φθ , *0 1 0 0 1 1[ ,..., , , ,..., ]Tp p m p n nk z k z k p p− − −= λ − λ −θ ,
1 2( ) [ ( ) , ( ) ]T T Tmt t tφ = ω ωC , 1 ( 1) ( 1)[ , ]m m m n m+ + × − −=C I 0 ,
1 12 2
( ) ( ) ( ),( ) ( ) ( ).t t u tt t y t
λλ
ω = ω +ω = ω +
A bA b (15)
54
Normalized gradient algorithm with projection (assumption 2):
( ) ( ) ( )t t t= +θ g f , 0(0) =θ θ , 0t ≥ , (16)
2
( )( ) ( )
( )
tt t
m t
φ= −εg Γ , ( ) ( ) ( ) ( )Tt t t y tε = φ −θ , ( ) 1 ( ) ( )Tm t t t= + κφ φ ,
0 if ( ) or( ) and ( ) 0 or( )( ) and ( ) 0,
( ) otherwise,
k k kk k k
kk k k
k
tt g tf tt g t
g t
⎧ θ < θ < θ⎪ θ = θ ≥⎪= ⎨ θ = θ ≤⎪
−⎪⎩
, 1,k nθ= .
Properties:
( )tθ , ( )tθ , ( ) / ( )t m t L∞ε ∈ and ( )tθ , 2( ) / ( )t m t Lε ∈ .
2) Reference model selection: ( )[ ]( ) ( )m mP s y t r t= , (17)
where ( )mP s is a stable polynomial of degree n m− and ( )r t is a bounded and piecewise
continuous reference input signal.
55
3) Controller structure construction:
1 1 2 2 3 4( ) ( ) ( ) ( ) ( )c T c c T c c cu t t t y t r t= ω θ + ω θ + θ + θ , (18)
where 1c nθ ∈ , 2
c nθ ∈ , 3cθ ∈ , 4
cθ ∈ are the controller parameters,
1( )( ) [ ]( )( )
c
c
a st u ts
ω =Λ
, 2( )( ) [ ]( )( )
c
c
a st y ts
ω =Λ
, 2( ) [1, ,..., ]n Ta s s s −= ,
and 1 22 1 0( ) ...n c n c c
c ns s s s− −−Λ = + λ + + λ + λ is a stable polynomial. A variant of realization:
1 1
2 2
( ) ( ) ( ),( ) ( ) ( ),
c c c c
c c c ct t u tt t y t
λ
λ
ω = ω +ω = ω +
A bA b
0 1 3 2
0 1 0 0 00 0 1 0 0
0 0 0 1c
c c c cn n
λ
− −
⎡ ⎤⎢ ⎥⎢ ⎥= ⎢ ⎥⎢ ⎥⎢ ⎥−λ −λ −λ −λ⎣ ⎦
A ,
0
01
c⎡ ⎤⎢ ⎥
= ⎢ ⎥⎢ ⎥⎣ ⎦
b .
( )y t( )u t( )r t
( )d t
1 ( )c tω
2 ( )c tω
( )[ ]( ) ( )[ ]( ) ( )pP s y t k Z s u t d t= +
1cθ
2cθ
3cθ
4cθ
( )( )c
a ssΛ
( )( )c
a ssΛ
≈ (15)
56
The controller parameter equation:
1 2 3 4( ) ( ) [ ( ) ( )] ( ) ( )[ ( ) ( ) ( )]T c T c c cc p c p ma s P s a s s k Z s s P s k Z s P sθ + θ + θ Λ = Λ − θ . (19)
Multiply (19) on ( )y t and substitute (14) for the case ( ) 0d t = :
1 2 3
4
( ) ( )[ ]( ) [ ( ) ( )] ( )[ ]( )
( ) ( )[ ]( ) ( ) ( ) ( )[ ]( ),
T c T c cc p
cc p c m
a s P s y t a s s k Z s y t
s P s y t k s Z s P s y t
θ + θ + θ Λ =
= Λ − θ Λ ⇒
1 2 3
4
( ) ( )[ ]( ) [ ( ) ( )] ( )[ ]( )
( ) ( )[ ]( ) ( ) ( ) ( )[ ]( ).
T c T c cp c p
cc p p c m
a s k Z s u t a s s k Z s y t
s k Z s u t k s Z s P s y t
θ + θ + θ Λ =
= Λ − θ Λ
Now divide both sides on ( ) ( )c ps k Z sΛ ( ( )Z s and ( )c sΛ are stable polynomials):
2 314
( ) ( )( ) [ ]( ) [ ]( ) ( ) ( )[ ]( ),( ) ( )
T c cT ccc
mc c
a s sa s u t y t u t P s y ts s
θ + θ Λθ + = − θΛ Λ
⇒
1 1 4 2 3 4( ) ( ) ( ) ( ) ( )[ ]( )c T c c T c c cmt t y t u t P s y tω θ + ω θ + θ = − θ .
Substitution of the control (18) gives
4 4( )[ ]( ) ( )c cmP s y t r tθ = θ ⇒ 4 4( )[ ]( ) ( )[ ]( )c c
m m mP s y t P s y tθ = θ ⇒ ( )[ ]( ) [ ]( ) 0m mP s y t y t− = .
57
4) Controller parameter calculation: 1
4c
pk−θ = ⇒ ( ) ( )[ ( ) ( ) ( )]c mB s s P s Z s P s= Λ − , then (19) takes the form:
1 2 3( ) ( ) [ ( ) ( )] ( ) ( )T c T c cc pa s P s a s s k Z s B sθ + θ + θ Λ = .
The right hand side is a polynomial of degree 2 2n − with coefficients linearly dependent on
1cθ , 2
cθ and 3cθ . The left hand side is a polynomial of degree 2 2n − with constant coefficients.
Equating the coefficients with the same powers of s we obtain the solution:
1 1 1 0 1 0 2 0( ,..., ; ,..., ; ,..., )cn m np p z z− − −θ = Θ λ λ , 2 1 1 0( ,..., )c
np −θ = Θ λ , 3 1 1 0( ,..., )cnp −θ = Θ λ
*
1 1 2 0( ; ,..., )cn−θ = Θ λ λθ , *
2 1 2 0( ; ,..., )cn−θ = Θ λ λθ , *
3 1 2 0( ; ,..., )cn−θ = Θ λ λθ .
Example 1: 11 1 2(1 )c
ma−θ = θ − θ − , 12 1c
mb−θ = θ .
Theorem 1. Under assumption 2 and that all zeros of ( )Z s are stable:
(i) ( )y t , ( )tθ , ( )tθ , 1( )tω , 2( )t L∞ω ∈ ; (ii) 2( ) ( )my t y t L− ∈ , lim [ ( ) ( )] 0t my t y t→∞ − = .
58
b. Pole placement control The pole placement equation:
*( ) ( ) ( ) ( ) ( ) ( )A s C s Q s P s D s Z s= + , (20)
where *( )A s is the desired polynomial of the closed loop system; ( )C s and ( )D s are
polynomials of the pole placement control:
1 1( ) ( ) ( ) ( ) ( )[ ]( ) ( ) ( )[ ]( )c c cu t s C s Q s s u t D s s r y t− −= Λ − Λ + Λ − , (21)
where ( )r t is a bounded and piecewise continuous reference input signal, ( )[ ]( ) 0Q s r t = ⇒
(a) ( ) 0r t = ⇒ ( ) 1Q s = ; (b) ( ) 0r t c= ≠ ⇒ ( )Q s s= ; (c) ( ) atr t ce−= ⇒ ( )Q s s a= + , 0a > .
According to (21) the control is a dynamical system:
( ) ( )[ ]( ) ( )[ ]( )C s Q s u t D s r y t= − . (22)
Controller structure ( 21( ) [1, ,..., ]n n Ta s s s θ + −= ):
1 1 2 1 3( ) ( ) ( )[ ]( ) ( ) ( )[ ]( ) ( ) ( )cT cT cc cu t a s s u t a s s y r t y t r t= θ Λ + θ Λ − + θ − .
59
Properties:
1) multiplying both sides of (20) on ( )y t we obtain:
*( )[ ]( ) ( ) ( ) ( )[ ]( ) ( ) ( )[ ]( )( ) ( ) ( )[ ]( ) ( ) ( )[ ]( ) ( ) ( )[ ]( )( ) ( )[ ]( ).
A s y t C s Q s P s y t D s Z s y tC s Q s P s y t Z s D s r t C s Q s u tZ s D s r t
= + == + − ==
(23)
( )r t L∞∈ and *A is stable ⇒ ( )y t L∞∈ . 2) multiplying both sides of (20) on ( )u t we obtain:
*( )[ ]( ) ( ) ( ) ( )[ ]( ) ( ) ( )[ ]( )
( ) ( )[ ]( ) ( ) ( )[ ]( ) ( ) ( )[ ]( ).A s u t C s Q s P s u t D s Z s u t
P s D s r y t D s Z s u t P s D s r t= + == − + =
( )r t L∞∈ and *A is stable ⇒ ( )u t L∞∈ .
3) using (20)−(23) we get:
*( )[ ]( ) 0A s y r t− = ⇒ lim ( ) ( ) 0t y t r t→∞ − = .
Assumption 3. ( ) ( )Q s P s and ( )Z s are coprime.
Theorem 2. Under assumption 3 all signals are bounded and lim [ ( ) ( )] 0t y t r t→∞ − = .
60
SUMMARY
Direct adaptive control Indirect adaptive control
Structure + −
Parameterization − +
Restrictions (minimum phase)
Certainty equivalence
Pm(θ* )
u y C(θс)
Adaptive parameter estimation
Control parameter derivation
( )tθ
θc(t)
Pm( *cθ )
u y C(θс)
Adaptive parameter adjustment
θc(t)
61
Example 1
Indirect adaptive control ⇔ Robust control
Plant: y ay bu d= − + + .
Assumption: 0 a a< ≤ , 0 b b< ≤ + 0ma > , ( ) 0r t = .
Normalized gradient descent algorithm with projection.
Robust control: u ky= , 1min ( ),0mk b a a−= − .
1.5a = , 2b = , 0.5a = , 0.1b = , 5ma = , ( ) 0d t = , ( ) 0v t = .
0 0.02 0.04 0.06 0.08 0.1
0.2
0.4
0.6
0.8
1
ya
yr
ym
t
0 0.02 0.04 0.06 0.08 0.1
50−
40−
30−
20−
10−ua
ur
t
62
1ma = , ( ) 0d t = , ( ) 0v t = .
0 0.2 0.4 0.6 0.8
0.2
0.4
0.6
0.8
1
ya
yr
ym
t
0 0.2 0.4 0.6 0.8
10−
8−
6−
4−
2−ua
ur
t
1ma = , ( ) 5sin(5 )d t t= , ( ) 0v t = .
0 1 2 3 4 5
0.5−
0.5
1
ya
yr
ym
t
0 1 2 3 4 5
10−
5−
5
ua
ur
t
1ma = , ( ) 0d t = , ( ) 0.1sin( )v t t= .
0 1 2 3 4 5
0.5−
0.5
1
ya
yr
ym
t
0 1 2 3 4 5
10−
8−
6−
4−
2−
2
ua
ur
t
63
5. ADAPTIVE OBSERVERS
A nonlinear system in state space presentation:
( ) ( )= + +x Ax B y u φ y , =y Cx , (24)
n∈x , m∈u , p∈y are the state, the input (control) and the measurable output;
A , C are constant and known, the functions ( )B y and ( )φ y are continuous and known.
Everything is known except the state x (it is not measurable) ⇒ the state observer design:
( ) ( ) [ ]= + + + −x Ax B y u φ y L y Cx ,
x is the estimate of x ; L is the observer matrix gain, −A LC is Hurwitz.
Assumption 1. ( )t L∞∈x , ( )t L∞∈u for all 0t ≥ .
The estimation error = −e x x :
( ) ( ) ( ) ( ) [ ] [ ]= − = + + − + + + − = −e x x Ax B y u φ y Ax B y u φ y L y Cx A LC e .
The matrix −A LC is Hurwitz (design of L ) ⇒ ( )t L∞∈x , lim [ ( ) ( )] 0t t t→∞ − =x x .
64
A nonlinear system with parametric uncertainty:
( ) ( ) ( , )= + + +x Ax B y u φ y G y u θ , =y Cx , (25)
q∈θ is the vector of unknown parameters, ( , )G y u is a known continuous function.
The adaptive observer:
( ) ( ) [ ] ( , )= + + + − + −x Ax B y u φ y L y Cx G y u θ Ωθ , (26)
[ ] ( , )= − −Ω A LC Ω G y u , (27)
[ ]T T= −γ −θ Ω C y Cx , 0γ > , (28)
q∈θ is the estimate of θ , n q×∈Ω is an auxiliary filter variable.
The state estimation error = −e x x :
[ ] ( , )[ ]= − + − +e A LC e G y u θ θ Ωθ .
−A LC is Hurwitz + Properties of ( )tθ and ( )tθ ⇒ Properties of ( )te .
65
The auxiliary error [ ]= + −δ e Ω θ θ :
[ ][ ] ( , )[ ] [ ] ( , )[ ] [ ] .
= + − − == − + − + + − − − − = −
δ e Ω θ θ ΩθA LC e G y u θ θ Ωθ A LC Ω G y u θ θ Ωθ A LC δ
−A LC is Hurwitz ⇒ ( )t L∞∈δ , lim ( ) 0t t→∞ =δ .
−A LC is Hurwitz + ( )t L∞∈y , ( )t L∞∈u (assumption 1) ⇒ ( )t L∞∈Ω .
The parameter estimation error ( ) ( )t t= −θ θ θ :
[ ] [ ]T T T T T T= − = γ − = γ = γ −θ θ Ω C y Cx Ω C Ce Ω C C δ Ωθ .
Intuition: lim ( ) 0t t→∞ =δ ⇒ ( ) ( )Tt t= −γθ h h θ , ( ) ( )T Tt t=h Ω C for 0t ≥ big enough.
Assumption 2. ( )th is PE: 0∃ρ > , 0δ > : 0 ( ) ( ) qt T
nd tτ τ τ ≥ ρ∫ h h I , t∀ ≥ δ .
Assumption 2 ⇒ ( )t L∞∈θ , lim ( ) 0t t→∞ =θ + properties of ( )tδ ⇒ ( )t L∞∈e , lim ( ) 0t t→∞ =e .
Theorem 1. Under assumptions 1 and 2 all signals in (25)−(28) are bounded and
lim [ ( ) ( )] 0t t t→∞ − =x x , lim [ ( ) ] 0t t→∞ − =θ θ .
66
Example 2
Oscillating pendulum:
2 sin( ) cos( ) ( ) ( )y y y b y f t d t= −ω − ρ + + , (13)
[ , )y = ϕ∈ −π π is the measured angle, y ∈ and y ∈ are the angle velocity and
acceleration; 0ρ > is an known friction coefficient, 0ω > is an unknown natural frequency,
0b > is an unknown control gain.
Presentation in the form (25) for 1x y= , 2x y= , u f= and ( ) 0d t = :
1 2 12
2 2 1 1
, ,sin( ) cos( ) ( ).
x x y xx x x b x u t
= == −ρ − ω +
⇒ 0 10⎡ ⎤= ⎢ ⎥−ρ⎣ ⎦
A , 10
T⎡ ⎤= ⎢ ⎥⎣ ⎦C , ( ) 0y =B , ( ) 0yϕ = ,
0 0( , ) sin( ) cos( )y u y y u⎡ ⎤= ⎢ ⎥−⎣ ⎦
G , 2
b⎡ ⎤ω= ⎢ ⎥⎣ ⎦
θ .
Both assumptions are satisfied for this example.
67
1ω = , 0.1ρ = , 0.5b = , ( ) sin(3 )f t t= , [2,1]T=L , 1000γ = .
0.3− 0.2− 0.1− 0 0.1 0.2 0.30.4−
0.2−
0
0.2
0.4
ϕ'
ϕ
0 2 4 6 8 100.5−
0
0.5
1
1.5
ω2
b
θ1
θ2
t
0.4− 0.2− 0 0.2 0.40.4−
0.2−
0
0.2
0.4
ϕ'
ϕ
0 2 4 6 8 100.5−
0
0.5
1
1.5
ω2
b
θ1
θ2
t
0.4− 0.2− 0 0.2 0.40.4−
0.2−
0
0.2
0.4
ϕ'
ϕ
0 2 4 6 8 100.5−
0
0.5
1
1.5
ω2
b
θ1
θ2
t
( ) 0d t =
( ) 0.5sin(10 )d t t=
( ) 0.5sin(6 )d t t=
68
INDIRECT ADAPTIVE CONTROL
OUTLINE
1. Introduction a. Main properties b. Running example
2. Adaptive parameter estimation a. Parameterized system model b. Linear parametric model c. Normalized gradient algorithm d. Normalized least-squares algorithm e. Discrete-time version of adaptive algorithms
3. Identification and robustness a. Parametric convergence and persistency of excitation b. Robustness of adaptive algorithms
4. Indirect adaptive control a. Model reference control b. Pole placement control
5. Adaptive observers
Recommended